API reference

Six endpoints under https://api.etal-labs.com/v1. Requests and responses are JSON.

Basics

Authentication

Send your key in the Authorization header as a bearer token. x-api-key is accepted too. Keys are created and revoked in the console, and a revoked key stops working at once.

POST /v1/detect HTTP/1.1
Host: api.etal-labs.com
Authorization: Bearer etal_...
Content-Type: application/json

Text format

All text is Markdown. Write math as $…$, citations as [@key] and footnotes as [^1]; QED is trained to carry them through unchanged. Separate paragraphs with a blank line.

Models and versions

A model string names a fixed checkpoint and does not change: qed-2026b, sic-2026a. The aliases qed and sic point at the current default of each family and move when a new version is released. Pin the dated string if you need results to stay comparable.

Limits

Each account has a rate limit in requests per minute, a cap on concurrent requests and a monthly allowance in US dollars at list price. A request that would exceed one of them is refused with status 429 and an error type that says which. Retry-After is set when waiting will help. Models and limits has the defaults.

What is stored

For each request we keep the endpoint, model, sizes, cost, latency and status. We do not keep the text you send or the text returned.

Write

QED writes a passage from notes (compose) or edits a draft (polish, revoice). It is not streamed; a 200-word passage takes about three seconds.

The response:

{
  "id": "wr_d5c89318-…", "object": "writing", "model": "qed-2026b", "task": "compose",
  "text": "Vitamin D did not reduce fractures (HR 0.98, 95% CI 0.89 to 1.07, 25,871 participants, median follow-up 5.3 y), even in the subgroup with low baseline levels.",
  "words": 27, "truncated": false,
  "guard": {"samples": 1, "repaired": false, "problems": []},
  "usage": {"input_tokens": 155, "output_tokens": 60, "cost_usd": 0.00067},
  "notes": []
}

Detect

Sic scores every sentence of a text. It takes about two seconds per 1,000 words.

The response:

Compare

Checks a model's output against its source. Use it on every QED output.

The response has warnings, each with a type and a severity, and pass, which is true when no warning is an error.

The judge is a language model and can be wrong. It quotes the passages it relies on, so each finding can be checked.

Disclosure

Drafts an AI-use disclosure from your account's own recent requests. It counts calls; it has no access to the text.

The response has statement, placement (where that style of journal expects it), based_on (the counts used) and authors_must_confirm. The statement reports only what your request log shows. Its closing sentence, that the authors reviewed and edited the output, is yours to make true.

Models and usage

Lists the models with their limits, list prices and status, and the models that are named but not released. A checkpoint that has been replaced has status superseded and stays callable by its dated string.

One model. Aliases resolve to the model they point at.

This month's allowance, the amount used, your rate limits, and counts per endpoint.

Errors

Errors come back with a matching HTTP status and one shape:

{"error": {"type": "input_too_long", "message": "`text` has 2500 words; the limit is 2000. Split the text by section and send the parts separately.", "param": "text"}}